Philippe Marschall wrote:
2008/9/27 Joshua Gargus <schwa@fastmail.us>:
  
Philippe Marschall wrote:

2008/9/26, Jason Johnson <jason.johnson.081@gmail.com>:


On Thu, Sep 25, 2008 at 9:15 PM, Igor Stasenko <siguctua@gmail.com> wrote:


2008/9/25 Chris Kassopulo <ckasso@sprynet.com>:


http://www.lesser-software.com/en/content/products/lswvst/lswvst.htm




It would be nice to look at it (LSWV).
So many cool features. Too bad its proprietary. :(

--
Best regards,
Igor Stasenko AKA sig.


Is the cost prohibitive?  Dolphin wasn't free, but it was really cheap
and a quality product.  Keep in mind that while programming languages
have been largely "commoditized"  the best implementations of many
languages still cost money.  E.g. you want the best C compiler?  It's
not GCC.  Buy the Intel compiler and watch your code run twice as
fast.




Let me preface this by saying that I support free software, and use it in
preference to proprietary software when possible.  However, I disagree with
much of what you say.

If you define best best as "longest bar in some benchmark".

I define best as "producing the best performance for the program that I am
interested in compiling".

Intel doesn't manage to charge serious money for their compiler because it
is better on some synthetic benchmark but worse in the real world.  Maybe
GCC has mostly caught up with 4.3, but there's no doubt that ICC has
traditionally generated faster code.  I don't have extensive/varied personal
experience with ICC, but if you compile Squeak with it on Intel, you'll see
a 20% speed-up on macro benchmarks.

If you
want to run your software only on Intel chips, no AMD, no Motorola, no
Sun, no ....

First, it's simply not true that Intel's compiler doesn't work for AMD.
After surfing around for 15min or so, I learned:
- at various times in the past, AMD chose ICC as the compiler they used to
generate SPEC benchmarks.
- there are numerous personal accounts where ICC generates the fastest code
for AMD for their pet application.

Typically, Intel CPUs have a higher percentage improvement by using ICC
instead of GCC, but AMD CPUs also benefit.
    

That's funny because AMD accuses Intel of checking for CPUID:

http://www.amd.com/us-en/assets/content_type/DownloadableAssets/AMD-Intel_Full_Complaint.pdf

page 40 and 41
  

Thanks for the link, I forgot about those allegations.  Now that I have the search term "amd cpuid intel compiler", I find plenty of instances of people claiming to have independently discovered CPUID checks in code generated by ICC.

Given some of the things I've read (ICC generating ridiculously slow memcpys for AMD), I wonder how ICC can still be competitive with top compilers in HPC benchmarks like:
http://www.cse.scitech.ac.uk/disco/Benchmarks/Opteron_compilers.pdf

It is easier to find anecdotes on less-official-looking pages like the following...
http://www.altechnative.net/e107_plugins/content/content.php?content.6
... where ICC does well.

Nevertheless, even though ICC does well on AMD, it could do better without the CPUID checks, and I hope that Intel had to pay AMD a lot of money for their actions.  What was the result of AMD's litigation?

Cheers,
Josh

Cheers
Philippe